After a year, BigScience’s AI language model is finally available – TechCrunch


After more than a year of planning and training, the volunteer-led project has created an open-source language model that they say is as powerful as OpenAI. GPT-3, but free and open to anyone (if they have the computing power). The model, called Bloom, is available as open source, along with the code and datasets used to create it. A Brooklyn AI startup Hugging face released a free web app that lets anyone try Bloom without having to download it.

Bloom is the brainchild of BigScience, an international, community-based project that aims to make large natural language models widely available for research. Large language modelsor “LLMs” for short, can translate, summarize and write text with a human touch; More or less?. (See GPT-3). But they have historically been expensive to create, keeping them away from researchers and firmly in the hands of big tech companies like Meta, Google and Microsoft.

That’s finally changing, thanks in part to the efforts of BigScience. The group’s more than 1,000 volunteer researchers, supported by ethicists, philosophers, legal scholars and engineers from startups and major tech companies, have been working for months on Bloom, which has competition from the likes of OpenAI and Alphabet’s DeepMind scaled up LLMs. One of the largest open source models for working with multiple languages, Bloom is designed for use in a variety of research applications, such as extracting information from historical texts.

Bloom is capable of generating text in 46 natural languages ​​and dialects and 13 programming languages,” according to a blog post with TechCrunch ahead of its release. “Although it has never been trained on any of those specific tasks, Bloom can be asked to create summaries or translations of text, output code from instructions, and follow instructions to complete original tasks such as writing recipes, extracting information from a news article, or writing. sentences using a newly defined made-up word … Bloom’s performance will continue to improve as the workshop continues to experiment and progress on top of Bloom.”

BigScience supporters also hope Bloom will spur new research on ways to combat the problems that plague all LLMs, including bias and toxicity. LLMs are trending remove fakes and exhibit prejudices against religions, genders, races and persons with disabilities. They also struggle with basic tenets of writing, often changing the subject without segue and endlessly repeating or even contradicting themselves.

“[Bloom] demonstrates the continued power of open source and open science even for expensive, large fundamental models,” You.com CEO and former Salesforce chief scientist Richard Socher told TechCrunch via email. Socher is not involved with BigScience. “It also shows that no one organization has a long-term lead in artificial intelligence. Once an organization shows that something can be done, the same opportunities will appear in other places six to 12 months later.”

Humble beginnings

BigScience’s origins lie in discussions years ago with Hugging Face’s Chief Scientific Officer, Thomas Wolfe. of GENCI Stefan Rekvena and IDRIS:Pierre-Francois Lavallee. The founders planned to create software, datasets, LLMs and tools to explore the social impact of AI, which has only received much attention from the research community in recent years.

Steering committees were soon formed to provide scientific and general advice to BigScience members from more than 60 countries and 250 institutions, design collaborative tasks, and organize workshops, hackathons, and public events. Various working groups have been tasked with addressing challenges such as data management, theorem proving of mathematics and archival strategies, as well as privacy and informed consent and other legal issues.

Bloom is the sum of their work. It was trained using $7 million worth of publicly funded (through grants) computing time on the Jean Zay supercomputer near Paris, France, which is among the most powerful machines in the world.

a hard is the discussion continuous in academic circles on the carbon impact of AI learning; data centers are not particularly environmentally friendly. But BigScience says Jean Zey, thanks to its unique cooling system and nuclear power source, was able to train Bloom to have a carbon footprint equivalent to a flight from Paris to New York.

Like all language models, Bloom is essentially a statistical tool for predicting words. Through a massive number of examples from a 1.6 terabyte training database, Bloom learned how likely words were to occur based on patterns, including the semantic context of the surrounding text. For example, given a typical letter that ends with “Looking forward…”, Bloom might complete it with “… to listening back”.

One of the goals of the BigScience working groups was to collect data representative enough to train Bloom. Due to systematic bias in public data sources, non-English LLMs have traditionally not performed as well as their English-language counterparts. Based on books, academic publications, radio transcripts, podcasts and websites, the 341 billion-word dataset used to train Bloom aims to encode different cultural contexts in different languages, including Swahili, Catalan, Bengali and Vietnamese.

The BigScience teams hand-picked nearly two-thirds of the dataset from 500 sources, with input from community groups including the African Natural Language Processing Community, Masakhane, LatinX in AI, and Machine Learning Tokyo. They edited for privacy and filtered for quality, for example to try to reduce the overrepresentation of porn sites that tend to carry sexist associations.

Bloom is not completely free of bias, no LLM is. But the hope is that by maintaining transparency around learning data, it will be easier for researchers to get to the roots of Bloom’s predictions and decision-making.

In large sizes

With 176 billion parameters, Bloom is roughly the size of GPT-3. Machine learning parameters are those parts of LLM that are learned from training data and tend to correlate with the model’s performance on a task such as text generation.

In general, models with higher parameters require more computing power for training. In 2020 study from AI21 Labs With just 1.5 billion parameters, the cost of developing a text generation model was set at $1.6 million; Bloom spent three months training on 384 Nvidia A100 GPUs. That fact has made it difficult for the community to use large, modern language models like Microsoft and Nvidia’s Megatron-Turing Natural Language Generation (MT-NLG), which has 530 billion parameters.

BigScience claims that rResearchers will be able to use Bloom for less than $40 an hour on a cloud provider. But is it?To overcome even this barrier to entry, the organization plans to release smaller, less hardware-intensive versions of Bloom and is developing a distributed system that will allow labs to share the model across their servers. The API works too.

Bloom joins a growing ecosystem of open source, highly capable LLMs with broad commercial and research applications. In February, the open AI research group EleutherAI released GPT-NeoX-20B, which at the time outperformed other public language models by several criteria. Months later, Meta open-sourced OPT-175B, which the company claims was the first model of the 175-billion-parameter language available to the AI ​​community.

They have been put to good use, businesses have already planted About EleutherAI models. But some researchers fear abuse. University of Maryland researchers found it possible for LLMs to create fake news and cyber security reports that persuasive enough to fool the pros. Another paper Co-authored by Meta researchers, the study examines the potential harm that can result from LLMs giving poor advice, particularly medical or psychological predictions.

Many companies that offer access to LLMs through an API, such as OpenAI, apply filters to remove problematic text. But open source models obviously have no such protection.

Recognizing the potential for abuse, Bloom provides documentation outlining its possibilities and limitations. Its use requires agreement to a legal license that obligates researchers not to use the model for malicious purposes. BigScience plans to monitor how the model is applied and adjust the license and documentation as necessary.

“We plan to add more languages, shrink the model so it’s easier to use with the same performance level, and we’ll support community efforts to expand it,” the blog post continues. “Bloom is a living family of models that will grow, not a one-off model.”



Source link

Leave a Reply

Your email address will not be published.